<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Kodownik</title>
    <description>The latest articles on DEV Community by Kodownik (@kodownik).</description>
    <link>https://dev.to/kodownik</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4216943%2F5aa2f725-b6d8-4c53-80a3-99946ee60f9a.png</url>
      <title>DEV Community: Kodownik</title>
      <link>https://dev.to/kodownik</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/kodownik"/>
    <language>en</language>
    <item>
      <title>One light-day: Voyager 1 crosses a line nothing we built has crossed before</title>
      <dc:creator>Kodownik</dc:creator>
      <pubDate>Tue, 17 Nov 2026 11:39:54 +0000</pubDate>
      <link>https://dev.to/kodownik/one-light-day-voyager-1-crosses-a-line-nothing-we-built-has-crossed-before-4g0f</link>
      <guid>https://dev.to/kodownik/one-light-day-voyager-1-crosses-a-line-nothing-we-built-has-crossed-before-4g0f</guid>
      <description>&lt;p&gt;Switch off the light for a moment. Stand by the window and imagine you turn on a torch and point it at the sky, towards the constellation of Ophiuchus.&lt;/p&gt;

&lt;p&gt;The beam passes the Moon in just over a second. After eight minutes it is as far away as the Sun. After four hours it crosses the orbit of Neptune, the last planet. And it keeps going.&lt;/p&gt;

&lt;p&gt;Meanwhile you have lunch. You answer emails. You go to bed. You wake up, make coffee, and the light is still travelling, all the time at the highest speed there is in the universe.&lt;/p&gt;

&lt;p&gt;Exactly one day later it reaches the place where Voyager 1 is on 18 November 2026 at 10:16 UTC.&lt;/p&gt;

&lt;p&gt;For the first time in history, something made by human hands is a full day of light-travel away from us.&lt;/p&gt;

&lt;h2&gt;
  
  
  A machine from the seventies
&lt;/h2&gt;

&lt;p&gt;Look at it up close. A dish antenna almost four metres wide. Below it, an octagonal box of 1970s electronics. On a long boom, a canister of plutonium that has been producing heat and power for half a century. The whole thing weighs less than a tonne.&lt;/p&gt;

&lt;p&gt;It launched on 5 September 1977. It flew past Jupiter, then Saturn. It moves at 17 kilometres per second. And still, to cover the distance light covers in a single day, it needed 49 years.&lt;/p&gt;

&lt;h2&gt;
  
  
  The bubble that isn't there
&lt;/h2&gt;

&lt;p&gt;Now imagine a bubble. A huge sphere around Earth, every point of which is exactly one day of light-travel away. That is 25.9 billion kilometres, 173 times farther than the Sun.&lt;/p&gt;

&lt;p&gt;I have to stop you here, honestly: the bubble doesn't exist. There is no wall out there, nothing breaks, the probe feels nothing. The boundary is a convention set by our clock. But what changes across it is very real.&lt;/p&gt;

&lt;p&gt;The conversation changes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two days for "how are you?"
&lt;/h2&gt;

&lt;p&gt;If you ask Voyager "how are you?", the question travels for a day. The answer comes back after another day. A short exchange takes two days.&lt;/p&gt;

&lt;p&gt;NASA's engineers know this rhythm well. In November 2023 Voyager 1 started sending gibberish: a single memory chip in its flight data computer had failed. The signal then took 22.5 hours one way, so every diagnostic attempt cost nearly two days of waiting. Over several months people on Earth rewrote software for a half-century-old machine, cutting code into pieces and moving it into memory that still worked. In April 2024 the probe was making sense again.&lt;/p&gt;

&lt;p&gt;Imagine fixing a computer where every click answers two days later.&lt;/p&gt;

&lt;h2&gt;
  
  
  A whisper from yesterday
&lt;/h2&gt;

&lt;p&gt;Voyager's transmitter runs on about 22 watts, roughly a fridge light bulb. That whisper spreads out over an enormous volume on its way. We hear it only because 70-metre dishes of the Deep Space Network are listening.&lt;/p&gt;

&lt;p&gt;And here is what stops me most. From 19 November, 02:18 UTC, every signal we receive from it has been travelling for more than a day. When we read its data, we are looking at yesterday.&lt;/p&gt;

&lt;h2&gt;
  
  
  The dot and the record
&lt;/h2&gt;

&lt;p&gt;In 1990 Voyager turned around one last time and photographed Earth from six billion kilometres. We fit in a dot smaller than a pixel. Today it is more than four times farther away.&lt;/p&gt;

&lt;p&gt;It carries the Golden Record: greetings in 55 languages, the sound of the sea, a heartbeat, Bach and Chuck Berry. In about 40,000 years it will pass the star Gliese 445 at a distance of 1.6 light-years. Nobody will see it.&lt;/p&gt;

&lt;p&gt;The plutonium in its power source weakens every year, and engineers keep switching instruments off so the rest can run. Most of them are already silent. I don't know how much longer we will hear it. Nobody knows exactly. It's a matter of years rather than decades.&lt;/p&gt;

&lt;p&gt;But it is still transmitting.&lt;/p&gt;

&lt;p&gt;Every "good morning" we send it today will arrive when, for us, it is already tomorrow.&lt;/p&gt;

&lt;h2&gt;
  
  
  How the film is made
&lt;/h2&gt;

&lt;p&gt;I built this moment as an 86-second film inside my browser simulation of the Solar System: &lt;strong&gt;&lt;a href="https://voyager.kodownik.co/" rel="noopener noreferrer"&gt;https://voyager.kodownik.co/&lt;/a&gt;&lt;/strong&gt; (sound is behind a button in the corner, browsers block autoplay audio).&lt;/p&gt;

&lt;p&gt;A few things that may interest you as a developer:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Real positions.&lt;/strong&gt; Voyager's trajectory comes from JPL Horizons state vectors, sampled daily and simplified in time to an error of at most 0.001 AU. The exact light-day moment comes from Horizons with one-minute resolution.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The camera is a pure function.&lt;/strong&gt; Each of the six shots is &lt;code&gt;pose(t, anchors) → {position, target, roll, fov}&lt;/code&gt;, where anchors (Sun, Earth, probe, membrane radius) come from the live scene. For the first 30% of every shot the pose blends from where the previous shot ended, so cuts never jump.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Smoothness is tested, not eyeballed.&lt;/strong&gt; A Node test steps the whole film at 60 Hz and fails if the view direction turns faster than 70°/s or its angular speed jumps between frames. It caught a real 155°/s whip on the approach to the probe (the camera was looking at a point 6 units ahead, then the target switched to a probe 4,000 units away).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Time-lapse with slow motion.&lt;/strong&gt; At scene scale nothing visibly moves at real speed (the hugely enlarged probe closes in on the boundary by about a third of its length per day), so the crossing shot runs a film clock of &lt;code&gt;a·sinh(k·(f − fc))&lt;/code&gt; hours: 2.6 days in 20 seconds, slowing down around the crossing, with the date and time-lapse factor on screen.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;No build step.&lt;/strong&gt; Three.js r128 from a CDN and plain ES2015 scripts in an IIFE dual-export pattern, so the same math file runs in the page and in &lt;code&gt;node --test&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Where the film simplifies: the probe and planets are enlarged (at true scale you would see nothing), and the glowing membrane is a visualisation of a clock, not a physical object.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This text was written with AI assistance (Claude); the facts and figures come from the sources listed below. The simulation and the film are my own work.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally written in Polish for the Kodownik reading room, with film stills and sources: &lt;a href="https://kodown1k.github.io/czytelnia/voyager-doba-swiatla/" rel="noopener noreferrer"&gt;https://kodown1k.github.io/czytelnia/voyager-doba-swiatla/&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Sources: NASA/JPL Horizons; NASA's announcement of the one-light-day milestone (via &lt;a href="https://earthsky.org/space/voyager-1-1-light-day-from-earth-november-17-18-2026/" rel="noopener noreferrer"&gt;EarthSky&lt;/a&gt; and &lt;a href="https://sciencenews.org/article/voyager-1-light-day-exploration" rel="noopener noreferrer"&gt;Science News&lt;/a&gt;); &lt;a href="https://spectrum.ieee.org/voyager-1" rel="noopener noreferrer"&gt;IEEE Spectrum&lt;/a&gt; on the 2023–2024 computer fix; &lt;a href="https://www.jpl.nasa.gov/images/pia17047-voyager-signal-spotted-by-earth-radio-telescopes/" rel="noopener noreferrer"&gt;JPL&lt;/a&gt; on the transmitter; &lt;a href="https://en.wikipedia.org/wiki/Gliese_445" rel="noopener noreferrer"&gt;Gliese 445 (Wikipedia)&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>space</category>
      <category>javascript</category>
      <category>threejs</category>
      <category>science</category>
    </item>
    <item>
      <title>In my own words</title>
      <dc:creator>Kodownik</dc:creator>
      <pubDate>Mon, 02 Nov 2026 12:27:35 +0000</pubDate>
      <link>https://dev.to/kodownik/in-my-own-words-l50</link>
      <guid>https://dev.to/kodownik/in-my-own-words-l50</guid>
      <description>&lt;p&gt;It's eight in the morning. The coffee is still brewing. In the team channel someone asks why the pipeline is red. You already know: the runner's disk is read-only, the code is fine. You ask your agent to reply.&lt;/p&gt;

&lt;p&gt;You get an essay. It opens with "I took a look." It has six sentences, four identifiers and the mechanism laid out step by step. It ends with "Let me know if you have any questions." Every word of it is true. None of it is you.&lt;/p&gt;

&lt;p&gt;You add three words: &lt;em&gt;in my words, concise, minimal&lt;/em&gt;. The second version is better. You send it. The next morning you do the same. And the next.&lt;/p&gt;

&lt;p&gt;Psychology knows a man who traced a star seen in a mirror day after day, each time sure he was doing it for the first time. After brain surgery Henry Molaison could no longer form new memories of facts and events, yet his hand still learned the task. Your agent also meets you for the first time every morning. Only this time, you're the one repeating the exercise.&lt;/p&gt;

&lt;p&gt;Those three words aren't a request. They're a memory that dies every morning.&lt;/p&gt;

&lt;h2&gt;
  
  
  The hand knows, the head doesn't
&lt;/h2&gt;

&lt;p&gt;Try describing your style. You'll get something like "short, to the point, human." Everyone thinks that about themselves, and everyone reads those words differently. Your "short" is two sentences. Someone else's is a paragraph without an intro.&lt;/p&gt;

&lt;p&gt;It's not your fault. Writing style lives in the same drawer as riding a bike. Your hand knows how you write; your head can't recite it.&lt;/p&gt;

&lt;p&gt;But there's a side door. Recalling something from a blank page is hard. Recognising it when it's lying in front of you is much easier. The whole method rests on that: you don't ask yourself how you write. You lay cards on the table and look.&lt;/p&gt;

&lt;h2&gt;
  
  
  Twelve cards
&lt;/h2&gt;

&lt;p&gt;Here's how we did it at Kodownik. First, material: two years of messages from the team chat and a few dozen comments from the issue tracker. Bots and private matters were dropped right away. Then twelve messages went on the table, shuffled, each with one question: mine or not mine?&lt;/p&gt;

&lt;p&gt;You'd expect a simple sort. Something else happened. Among the "mine" were messages written by someone else. Among the "not mine" landed four of my own.&lt;/p&gt;

&lt;p&gt;The question was "mine or not mine", but the answer turned out to be an answer to a different one: would I sign this?&lt;/p&gt;

&lt;p&gt;The accepted ones shared a spine: the cause in one sentence, what's been done, and the next step with a named owner. The rejected ones fell into two families. Some were overloaded. Others said too little, and the shortest had two words and a question mark: "Anyone on this?"&lt;/p&gt;

&lt;p&gt;Your style isn't how you write. It's how you write on your best day.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two versions side by side
&lt;/h2&gt;

&lt;p&gt;The details are settled with the same trick. For every doubt, two versions of the same message go on the table. The question isn't "which do you prefer", it's "which one do you send".&lt;/p&gt;

&lt;p&gt;Options as prose or as a list? A list, and when there's one clear path, a note saying "this is the one I'd pick". To someone who decides rather than codes? The effect instead of the mechanism, and one request for a decision at the end. To a tester? What to check and what the fix doesn't change. Nine such pairs, nine decisions. None required describing myself.&lt;/p&gt;

&lt;h2&gt;
  
  
  The hand instead of the head
&lt;/h2&gt;

&lt;p&gt;For the model, a style file is context, not a constraint. In Claude Code a skill is picked up when the model recognises a matching situation from the skill's description, and sometimes it won't. So next to it stands a gate: a &lt;code&gt;PreToolUse&lt;/code&gt; hook that won't let a message out to the chat or the issue tracker until the agent has read the style. The gate doesn't ask the model's opinion. (As of the November 2026 docs, a &lt;code&gt;PreToolUse&lt;/code&gt; hook can deny any tool call, MCP tools included.)&lt;/p&gt;

&lt;p&gt;Before it was saved, the file went through a test: three real situations from the previous days, rewritten in the new style. All three passed without a single correction. Had any needed one, each correction would have been a missing rule.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where this story overreaches
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;A style file isn't you.&lt;/strong&gt; The model imitates the shape of a message. The voice will neatly wrap a wrong diagnosis, too. The content and the responsibility are still yours.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Twelve cards is not much.&lt;/strong&gt; It's the calibration of one person in one conversation. Some rules are guesses, and the file labels them as such.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The gate checks that the style was read, not that it was followed.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Your turn
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Gather traces.&lt;/strong&gt; Messages to people from recent years. Drop the bots and private matters.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lay out twelve cards.&lt;/strong&gt; Mix yours with other people's. Answer only: mine or not mine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Choose from pairs.&lt;/strong&gt; Turn every doubt into two versions of the same message.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Write the portrait.&lt;/strong&gt; One model sentence, a skeleton, the voice, a "never" list, rules per audience, "not like this → like this" examples. Separately: where each rule came from and which ones are guesses.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Let it write a letter.&lt;/strong&gt; Three real situations from the past week. Every correction you make is a missing rule.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Put up the gate.&lt;/strong&gt; A hook before sending that refuses until the style has been read.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Leave a door.&lt;/strong&gt; A table of decisions with dates, kept in git.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;To start, you can paste this to your agent:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;I want to capture my style of writing to people as a skill. Walk me through a calibration:
1. Take my messages from [where] from the last two years, without bots and private matters.
2. Show me twelve, mixed with other people's. For each, ask only: mine or not mine.
3. From the accepted and rejected ones, extract a common skeleton. Show every doubt as two versions of the same message and let me choose.
4. Write SKILL.md (model sentence, skeleton, voice, never, audiences, "not like this → like this" examples) and SOURCES.md (where the rules came from, which are guesses).
5. Rewrite three of my real messages from the last week in the new style. Turn each of my corrections into a rule.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Henry Molaison could never say that he had practised the star. You can't say how you write. You don't have to. It's enough to recognise yourself on a card once.&lt;/p&gt;

&lt;p&gt;And from tomorrow on, you won't have to tell your agent who you are.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This text was written with AI assistance (Claude).&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally written in Polish for the Kodownik reading room: &lt;a href="https://kodown1k.github.io/czytelnia/moimi-slowami/" rel="noopener noreferrer"&gt;https://kodown1k.github.io/czytelnia/moimi-slowami/&lt;/a&gt; — with illustrations and the first part of the story, "The jay, the nutcracker and CLAUDE.md".&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Sources: Claude Code docs on &lt;a href="https://code.claude.com/docs/en/skills" rel="noopener noreferrer"&gt;skills&lt;/a&gt; and &lt;a href="https://code.claude.com/docs/en/hooks" rel="noopener noreferrer"&gt;hooks&lt;/a&gt;; &lt;a href="https://en.wikipedia.org/wiki/Henry_Molaison" rel="noopener noreferrer"&gt;Henry Molaison (Wikipedia)&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>writing</category>
      <category>productivity</category>
    </item>
    <item>
      <title>The jay, the nutcracker and CLAUDE.md</title>
      <dc:creator>Kodownik</dc:creator>
      <pubDate>Fri, 30 Oct 2026 10:33:24 +0000</pubDate>
      <link>https://dev.to/kodownik/the-jay-the-nutcracker-and-claudemd-ako</link>
      <guid>https://dev.to/kodownik/the-jay-the-nutcracker-and-claudemd-ako</guid>
      <description>&lt;p&gt;Picture a lab in Montreal. A man is tracing a star with a pencil, but he can see his hand only in a mirror. On day one the pencil wanders everywhere. By day three it barely slips. And every time, he is certain he has never done this before.&lt;/p&gt;

&lt;p&gt;That man is Henry Molaison, H.M. In 1953 surgeons removed the front parts of both his hippocampi to stop his epilepsy. The seizures stopped. So did his ability to form new lasting memories of facts and events. But his hand kept learning, something his mind never knew it had practised.&lt;/p&gt;

&lt;p&gt;You know this picture from your terminal. Every morning, in a fresh context window, your assistant sees the repo for the first time. Yesterday's two-hour talk about why the integration tests need a local Redis? Gone.&lt;/p&gt;

&lt;p&gt;So what can a brain, a jay and a nutcracker teach us about memory across sessions? Lesson one: "memory" isn't one device. It's a label on a dozen of them. (The Claude Code translations below are mostly proposals, not features; where I describe what exists, I rely on the October 2026 docs.)&lt;/p&gt;

&lt;h2&gt;
  
  
  Twelve letters and a sieve
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;George Sperling flashes twelve letters for a fraction of a second. You name three, maybe four. Yet if a tone right after the flash points at one row, you read it almost perfectly. All the letters were there. They just faded faster than you could say them. Sensory memory isn't a store. It's a sieve.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;In Claude Code the sieve is raw tool output: logs, stack traces, API responses. Compaction clears old tool results first, and that's fine: a file on disk can be read again. So nothing raw goes into memory. A note keeps the conclusion ("test X is flaky because it depends on the timezone"), not the log.&lt;/p&gt;

&lt;h2&gt;
  
  
  A workbench for about four things
&lt;/h2&gt;

&lt;p&gt;Now sit at a workbench. It's small. Nelson Cowan counted that it holds about four things at once, not the seven of the old saying. The only cheat is packing small things into one bundle you understand.&lt;/p&gt;

&lt;p&gt;For an agent, the workbench is the context window. MemGPT's researchers had the model raise the alarm at seventy percent full, to save what mattered before it was pushed out. Keep task state in a file that survives compaction, with a tabu list of "what I checked and ruled out", like a rat that knows which maze arms it has visited. And send big reading jobs to subagents: they return one bundle instead of fifteen facts. That's chunking.&lt;/p&gt;

&lt;h2&gt;
  
  
  The jay and the larva
&lt;/h2&gt;

&lt;p&gt;The western scrub jay buries food: a larva that spoils fast, a nut that keeps. In 1998 Nicola Clayton and Anthony Dickinson showed that the jay remembers what it hid, where and when. If the larva has waited too long, the jay doesn't go back for it. It goes for the nut. The researchers carefully called this &lt;em&gt;episodic-like&lt;/em&gt; memory, since nobody can show a bird relives anything. Hold on to that caution.&lt;/p&gt;

&lt;p&gt;"The build is green" spoils within an hour. "Migrations live in &lt;code&gt;migrations/&lt;/code&gt;" keeps for months. Claude Code stamps memory files with when they were written. What's missing is "good until": give each episode a shelf life, use absolute dates ("next Thursday" means nothing in December), and keep a log of what didn't work. Zep stores episodes losslessly, yet by vendor-reported numbers it scores 17.7% worse on questions about what the assistant itself said.&lt;/p&gt;

&lt;h2&gt;
  
  
  Knowledge without a memory
&lt;/h2&gt;

&lt;p&gt;You know Warsaw sits on the Vistula. You don't remember the lesson. That's semantic memory: knowledge that forgot where it came from.&lt;/p&gt;

&lt;p&gt;Frederic Bartlett had English listeners retell a Native American legend. "Canoes" became "boats"; whatever didn't fit vanished or got straightened. So the most precious lines in project notes are the exceptions, and they're the first to go.&lt;/p&gt;

&lt;p&gt;In Claude Code, semantic memory is CLAUDE.md, and the docs' criterion is pure cognitive science: add an entry when Claude repeats a mistake. Write the &lt;em&gt;why&lt;/em&gt;, not just the &lt;em&gt;what&lt;/em&gt;. When a decision changes, don't delete the entry; mark it invalid as of today. Newer wins, history stays.&lt;/p&gt;

&lt;h2&gt;
  
  
  The hand knows better
&lt;/h2&gt;

&lt;p&gt;Back to the mirror. H.M.'s hand learned with no memory of practice. That's procedural memory, and for agents it pays best. Voyager, a Minecraft agent with a library of skills stored as code, collected 3.3 times more unique items.&lt;/p&gt;

&lt;p&gt;Claude Code splits this in two. Recipes go into skills, loaded when needed. What must happen every time goes into hooks, which run at fixed points no matter what the model decides. The docs say it plainly: CLAUDE.md is context, not configuration. The model reads it. It isn't obliged to obey.&lt;/p&gt;

&lt;p&gt;A hook is the hand tracing the star with no consciousness involved. So sort rules by the cost of breaking them: must never fail, a hook; needs judgement, CLAUDE.md; long recipe, a skill.&lt;/p&gt;

&lt;h2&gt;
  
  
  The library on your way home
&lt;/h2&gt;

&lt;p&gt;An overdue book sits in your bag. You forget it all day, until you walk past the library. That's prospective memory: memory for what you still have to do. Its enemy is interruption: in one study, two out of a hundred interrupted people thought about their intention during the break. For an agent, the break is the end of a session.&lt;/p&gt;

&lt;p&gt;Here's the surprise. The research review behind this piece found no agent system that evaluates prospective memory as its own kind. Yet Claude Code has the mechanism: a rule with a &lt;code&gt;paths:&lt;/code&gt; field loads exactly when Claude touches a matching file. The path is the library. Write intentions as "when you touch X, do Y", and delete the rule in the commit that does the job.&lt;/p&gt;

&lt;h2&gt;
  
  
  Moved sticks
&lt;/h2&gt;

&lt;p&gt;The nutcracker hides over thirty thousand seeds and finds them up to nine months later. Researchers shifted the landmarks by about twenty centimetres, and the bird dug with exactly the same shift. It remembers places relative to the sticks.&lt;/p&gt;

&lt;p&gt;Auto memory is a nutcracker's estate: &lt;code&gt;MEMORY.md&lt;/code&gt; is the cache map, topic files are the seeds. At startup only the first 200 lines or 25 KB of the index load. Past that line nothing is "remembered faintly". It simply isn't there. Put important lines on top, pin facts to a path and a symbol, and after a refactor grep for orphaned anchors.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sleep that reviews the day
&lt;/h2&gt;

&lt;p&gt;You fall asleep, and the hippocampus replays the day, weaving it slowly into old knowledge, because writing it in directly would smash what's there. But sleep also adds false "gist": meanings nobody said.&lt;/p&gt;

&lt;p&gt;Claude Code has the first pieces: &lt;code&gt;/doctor prompt-audit&lt;/code&gt; hunts contradictions and stale entries. What's missing is replay. Imagine a nightly timer running a consolidation skill on a separate branch: for each fresh episode, add, update, invalidate or nothing, and a diff waiting for morning approval. A proposal, not a write, because sleep invents gist too. And sleep-time compute research found only moderate savings on code. Do it for hygiene, not cheaper tokens.&lt;/p&gt;

&lt;h2&gt;
  
  
  A memory open for editing
&lt;/h2&gt;

&lt;p&gt;In 2000 Karim Nader showed that a recalled memory briefly turns unstable; a drug given right after recall erased it, six hours later it didn't. Elizabeth Loftus showed that information after an event contaminates memory; in "lost in the mall" studies about a third of people formed a partly or fully false one.&lt;/p&gt;

&lt;p&gt;A note the model rewrites on every read drifts like a story told twenty times. So edit an entry only after checking it in code, keep raw episodes untouchable, and keep memory in git, where &lt;code&gt;git diff&lt;/code&gt; shows reconsolidation in black and white.&lt;/p&gt;

&lt;h2&gt;
  
  
  The matriarch and the fruit fly
&lt;/h2&gt;

&lt;p&gt;The fruit fly forgets on purpose: the dopamine that writes a trace can erase it. Forgetting is a filter. Yet in 1993, when drought hit a park in Tanzania, herds led by old matriarchs fared better, an edge credited to memories of a drought over thirty years earlier.&lt;/p&gt;

&lt;p&gt;File memory never fades. It's cut off hard. So let entries expire unless they've proved their worth, archive instead of deleting, and give runbooks for rare failures matriarch status: they never expire.&lt;/p&gt;

&lt;h2&gt;
  
  
  The jay that was watched
&lt;/h2&gt;

&lt;p&gt;When a jay caches food while another jay watches, it later moves the cache, but only if it has stolen itself. It even remembers who watched. It keeps a ledger of sources.&lt;/p&gt;

&lt;p&gt;An entry from a web page, an issue or a dependency's README is a cache someone watched. In the MINJA study, malicious records were planted in an agent's memory through queries alone, in 98.2% of attempts. So every entry should know its source and confidence. Trust order: human, agent, tool, and tool-sourced entries wait in quarantine.&lt;/p&gt;

&lt;h2&gt;
  
  
  The ant doesn't remember
&lt;/h2&gt;

&lt;p&gt;A harvester ant colony lives twenty, thirty years; a worker, one at most. The colony remembers through traces left in its surroundings.&lt;/p&gt;

&lt;p&gt;Your subagents are workers and the repo is the colony. But give seven subagents their own worktrees and a status file exists in seven copies: seven truths. Ants never edit another's pheromone. So keep status outside the worktrees, say in the shared &lt;code&gt;.git&lt;/code&gt; directory, one file per agent. A &lt;code&gt;SubagentStop&lt;/code&gt; hook won't let an agent leave without a trace, and entries unrefreshed for a day evaporate into "stale".&lt;/p&gt;

&lt;h2&gt;
  
  
  Where these analogies lie
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;One workbench.&lt;/strong&gt; A brain keeps its drawers in different structures. CLAUDE.md, auto memory and rules all land in one context window and compete; only hooks are truly separate.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;"Episodic" is a metaphor.&lt;/strong&gt; A dated note has a "when" but no "I". Say "-like", as the jay researchers did.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Designed, not inherited.&lt;/strong&gt; Limits quoted here come from the October 2026 docs. Check before you build.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Biology argues with itself.&lt;/strong&gt; Consolidation theories still clash.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Vendor numbers.&lt;/strong&gt; Mem0, Zep and Letta dispute each other's benchmarks, none measured on programming tasks.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Credit where due: the docs admit text memory doesn't guarantee obedience, and point to what does.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cheat sheet
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Memory&lt;/th&gt;
&lt;th&gt;In nature&lt;/th&gt;
&lt;th&gt;Claude Code today&lt;/th&gt;
&lt;th&gt;Proposal&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Sensory&lt;/td&gt;
&lt;td&gt;sieve&lt;/td&gt;
&lt;td&gt;tool output&lt;/td&gt;
&lt;td&gt;nothing raw&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Working&lt;/td&gt;
&lt;td&gt;~4 things&lt;/td&gt;
&lt;td&gt;context window&lt;/td&gt;
&lt;td&gt;state file, tabu list&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Episodic-like&lt;/td&gt;
&lt;td&gt;jay&lt;/td&gt;
&lt;td&gt;auto memory, timestamps&lt;/td&gt;
&lt;td&gt;shelf life, failure log&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Semantic&lt;/td&gt;
&lt;td&gt;knowledge without origin&lt;/td&gt;
&lt;td&gt;CLAUDE.md, rules&lt;/td&gt;
&lt;td&gt;"why", invalidate&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Procedural&lt;/td&gt;
&lt;td&gt;H.M.'s hand&lt;/td&gt;
&lt;td&gt;skills, hooks&lt;/td&gt;
&lt;td&gt;costly rules as hooks&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Prospective&lt;/td&gt;
&lt;td&gt;library&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;paths:&lt;/code&gt; rules&lt;/td&gt;
&lt;td&gt;"touch X, do Y"&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Spatial&lt;/td&gt;
&lt;td&gt;nutcracker&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;MEMORY.md&lt;/code&gt;, 200 lines&lt;/td&gt;
&lt;td&gt;anchors, top first&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Consolidation&lt;/td&gt;
&lt;td&gt;sleep&lt;/td&gt;
&lt;td&gt;&lt;code&gt;/doctor prompt-audit&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;nightly diff to approve&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Reconsolidation&lt;/td&gt;
&lt;td&gt;recall reopens&lt;/td&gt;
&lt;td&gt;markdown&lt;/td&gt;
&lt;td&gt;verify first, git&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Forgetting&lt;/td&gt;
&lt;td&gt;fly, elephants&lt;/td&gt;
&lt;td&gt;hard cut&lt;/td&gt;
&lt;td&gt;archive, matriarchs&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Metamemory&lt;/td&gt;
&lt;td&gt;watched jay&lt;/td&gt;
&lt;td&gt;none&lt;/td&gt;
&lt;td&gt;provenance, quarantine&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Collective&lt;/td&gt;
&lt;td&gt;ant colony&lt;/td&gt;
&lt;td&gt;worktrees&lt;/td&gt;
&lt;td&gt;status outside tree&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h2&gt;
  
  
  Tomorrow morning
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Check what loads.&lt;/strong&gt; Run &lt;code&gt;/context&lt;/code&gt;. Memory that doesn't load doesn't exist.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Move gates into hooks.&lt;/strong&gt; "No push without approval" is a &lt;code&gt;PreToolUse&lt;/code&gt; block, not a sentence.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Measure the workbench.&lt;/strong&gt; Main CLAUDE.md under 200 lines.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Task state to a file&lt;/strong&gt; that survives compaction.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Pin one intention to a path.&lt;/strong&gt;&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Schedule sleep.&lt;/strong&gt; Fifteen minutes a week, in a calendar.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Six ideas from fresh eyes
&lt;/h2&gt;

&lt;p&gt;A separate agent got the research review and one request: six mechanisms that aren't in it, no invented facts. It discarded its typical first thoughts, like a vector store. All six can be built today. None has been tested in practice.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Questioning the witness.&lt;/strong&gt; Teams remember who knows what (Daniel Wegner's transactive memory). &lt;code&gt;SessionEnd&lt;/code&gt; logs topic and session ID to &lt;code&gt;who-knows.tsv&lt;/code&gt;, and a skill resumes that session as a fork (&lt;code&gt;claude -p --resume &amp;lt;id&amp;gt; --fork-session&lt;/code&gt;) with one question: what did you learn about X, and why? It fails when transcripts are cleaned up, and the witness knows only that day's code.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. A thymus for rules.&lt;/strong&gt; The thymus kills cells that attack the body's own tissue. A candidate hook is replayed against old tool calls, counting how often it would have blocked successful sessions, and rejected above a threshold. If "self" includes a chronic bad habit, it rejects the very rule meant to fix it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. A bet before the test.&lt;/strong&gt; Dopamine neurons fire for surprise, not for the expected. A hook refuses to run tests without a bet, green or red, and only mismatches become memory candidates. The agent may learn to bet defensively, and nobody checked that tool results carry exit codes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Desire paths.&lt;/strong&gt; People wear shortcuts into grass; paving only the worn paths is an anecdote the agent admits it can't document. A hook logs which files successful sessions read together and proposes &lt;code&gt;paths:&lt;/code&gt; rules. Unfiltered, it also paves getting lost.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. A soil seed bank.&lt;/strong&gt; Seeds sleep for years and sprout after fire or ploughing. Archived episodes keep normalised error signatures; when a new error matches, a hook surfaces the old fix. Too coarse wakes false seeds, too fine wakes none, and messages change with library versions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;6. Mason's marks.&lt;/strong&gt; Medieval masons carved marks into stone, usually read as accounting for work, though interpretations vary; either way, the information travelled with the block. Commits carry &lt;code&gt;Why:&lt;/code&gt; and &lt;code&gt;Gotcha:&lt;/code&gt; trailers that a hook surfaces through &lt;code&gt;git blame&lt;/code&gt; before an edit, so memory follows the code through refactors. Squash merges and full rewrites break the chain, yet this is the one the agent would start with.&lt;/p&gt;

&lt;h2&gt;
  
  
  Coda
&lt;/h2&gt;

&lt;p&gt;H.M. helped science for half a century without remembering one of its thousands of studies. His hand remembered the star. Your agent has one advantage over him: someone can write it a note.&lt;/p&gt;

&lt;p&gt;The best memory knows how to forget. But it knows what it must never forget.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This text was written with AI assistance (Claude).&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Originally written in Polish for the Kodownik reading room: &lt;a href="https://kodown1k.github.io/czytelnia/sojka-orzechowka-i-claude-md/" rel="noopener noreferrer"&gt;https://kodown1k.github.io/czytelnia/sojka-orzechowka-i-claude-md/&lt;/a&gt; — with hand-drawn diagrams and the full source list.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Main sources: Claude Code docs on &lt;a href="https://code.claude.com/docs/en/memory" rel="noopener noreferrer"&gt;memory&lt;/a&gt; and &lt;a href="https://code.claude.com/docs/en/hooks" rel="noopener noreferrer"&gt;hooks&lt;/a&gt;; &lt;a href="https://arxiv.org/pdf/2009.01083" rel="noopener noreferrer"&gt;Memory systems of the brain (arXiv 2009.01083)&lt;/a&gt;; &lt;a href="https://memory.psych.missouri.edu/assets/doc/articles/2001/cowan-bbs-2001.pdf" rel="noopener noreferrer"&gt;Cowan 2001&lt;/a&gt;; &lt;a href="https://ai-papers.aktagon.com/papers/memgpt/" rel="noopener noreferrer"&gt;MemGPT&lt;/a&gt;; &lt;a href="https://en.wikipedia.org/wiki/Episodic-like_memory" rel="noopener noreferrer"&gt;Episodic-like memory&lt;/a&gt;; &lt;a href="https://ideas.repec.org/a/nat/nature/v406y2000i6797d10.1038_35021052.html" rel="noopener noreferrer"&gt;Nader et al. 2000&lt;/a&gt;; &lt;a href="https://arxiv.org/html/2503.03704v2" rel="noopener noreferrer"&gt;MINJA&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>ai</category>
      <category>productivity</category>
      <category>programming</category>
    </item>
  </channel>
</rss>
